Seminar Series

Fantastic DNimals and where to find them.

The introduction of deep neural networks has altered the fields of computer vision and machine learning. How DNNs relate to the field of biological vision is less clear. Are they functional or more mechanistic models of visual perception? I propose to consider DNNs in a similar way to how we consider model organisms, as systems that we can study to further understand how vision works, but with the possibility to inspect all units, being able to alter the methods and material with which DNNs are trained and even alter the architecture of these animals. Working from this perspective we have recently shown that training DNNs on unrelated tasks creates segregated representations while training on related task does not. This shows how architectural structure develops within DNNs for solving multiple tasks. This approach opens the way to training DNNs with multiple pathways solving multiple tasks, and relating these to human neuroimaging data. I will also present data which strongly indicates that, up to a certain degree, netwerk depth is a proxy for recurrent processing